| Description |
Machine learning is one of the main drivers of the hyper connected world. However, with billions of connected devices and millions of applications producing data at the edge, we can’t expect all the data to go to central cloud for processing, training and decision making in fastest and most cost-effective way. As a result, we need to be able to distribute the functions as much as possible.
Let’s for instance take the autonomous vehichle use case when cars need to constantly detect obstacles on the road and take actions accordingly. If the car has to send the information on a detected an obstacle back to the central cloud, the latency may be too large for the cars nearby to avoid the same obstacle. Using edge cloud, and V2V connectivity, nearby cars can be informed, and the information can at the same time be sent to central cloud to refine the model and then update all the vehicles.
Our goal is to use edge cloud to extend the reach and effectiveness of machine learning to communicate real-time data.
main objective:
The objective of this project is to develop a system where we can train the model on the cloud, send the model to the edge for the application to use and when the application identifies a new data semantic it takes the following actions:
- Inform the cars around it based on the proximity about the new information
- Let the training software in the cloud know of the new semantic
- The training engine to retrain the model
- The training engine to push the model to a subset of edge nodes (cars)also known as seeding
- The edge nodes can then automatically manage updating the network (each other) of the new model
During the project students have access to mimik team (business/technology) including the executives. Working on this projects not only enable you to work on latest technology i.e. edgeCloud but also provides opportunity for those that delivering the project successfully to potentially get hired by mimik as employee or Co-op.
Main deliverables:
The team should use mimik edge Cloud platform as the underlying communications platform. They could then utilize several mimik sample microservices (mModel) as the base and/or to learn about developing the model microservice on edge nodes using edgeSDK. They could use any cloud training model such as Google’s TensorFlow. The mimik cluster management enables car to car communication based on the scopes of network, proximity and account.
The main deliverable of this project is a “self-optimized model update” across all nodes. The research needs to answer the following important questions: What is the most optimum path for nodes to inform each other of the update? Given that the information (semantic) is regional and may be temporary (for instance, road blocks) should all cars in the city be informed of the model update? If not, are we going to face model fragmentation? What is the peer-to-peer data synchronization algorithm?
We expect a demonstration of this system at UBC as well as at the mimik office to all the mimik team members. The demonstration will be video recorded as well for the mimik YouTube Channel and demonstrations. The IP will belong to mimik, but team members’ names will be listed on the recorded video. Team members will have the opportunity to apply for job at mimik once their project is completed.
|